Brief

Kwai-Keye has released Keye-VL-2.0-30B-A3B, a new 30-billion parameter multimodal model designed for advanced video understanding and agent capabilities. The model excels in temporal localization, matching or surpassing Gemini-3-Flash on video benchmarks, and supports hour-long video contexts through its DSA-Native Long-Context Architecture. Keye-VL-2.0-30B-A3B also features a high-efficiency inference and training stack, robust post-training for reliable reasoning, and built-in agent abilities for tasks like code execution and tool use. AI

Kwai-Keye has released Keye-VL-2.0-30B-A3B, a new 30-billion parameter model designed for advanced long-video understanding and agent capabilities. The model demonstrates strong performance on video benchmarks, matching or exceeding Gemini-3-Flash in temporal grounding tasks. It features a DSA-native long-context architecture that allows for hour-long video comprehension with efficient computation, utilizing sparse attention and targeted feature aggregation. AI